For clustering a large Design Structure Matrix (DSM), computerized algorithms are necessary. A common algorithm by Thebeau uses stochastic hill-climbing to avoid local optima. The output of the algorithm is stochastic, and to be certain a very good clustering solution has been obtained, it may be necessary to run the algorithm thousands of times. To make this feasible in practice, the algorithm must be computationally efficient. Two algorithmic improvements are presented. Together they improve the quality of the results obtained and increase speed significantly for normal clustering problems. The proposed new algorithm is applied to a cordless handheld vacuum cleaner.